The agent pipeline
Infrastructure becomes a reasoning loop.
Specialized agents reason about intent while deterministic control-plane boundaries limit authority over state-changing actions.
01Intent AgentExtract constraints and objectives.
02Blueprint AgentCompile typed desired state.
03Arbitration AgentCompare provider candidates.
04Policy GateOPA rules + approvals.
05ExecutionControlled adapter.
06VerificationTest and reconcile.
Human in the loop
Agents propose. People authorize.
State-changing operations stop at explicit boundaries. A plan is a first-class artifact, not an implicit side effect of chat.
- Reviewable diff-style plans
- Approval gates before apply
- Policy denial independent of model output
- Auditable decisions and API calls
Continuous control
Verification is not the end.
Post-apply verification checks actual state, while reconciliation compares reality to the blueprint over time.
- Desired-state versioning
- Drift detection
- Policy-aware remediation
“A successful API call is not proof that the infrastructure is correct.”
AIaC design principle